A physically-constrained Gaussian spectral decomposition framework for remote sensing estimation of chlorophyll-a and suspended matter retrieval in highly turbid coastal waters

Titelangaben

Moradi, Masoud ; Lu, Meng ; van der Molen, Johan ; Arabi, Behnaz:
A physically-constrained Gaussian spectral decomposition framework for remote sensing estimation of chlorophyll-a and suspended matter retrieval in highly turbid coastal waters.
In: ISPRS Journal of Photogrammetry and Remote Sensing. Bd. 238 (2026) . - S. 314-336.
ISSN 0924-2716
DOI: https://doi.org/10.1016/j.isprsjprs.2026.05.011

Volltext

Link zum Volltext (externe URL):

Abstract

Retrieving water constituent concentrations (WCCs) from satellite observations remains challenging in optically complex and highly turbid coastal waters, where multiple constituents jointly control the spectral reflectance signal. This study develops and evaluates a physically Gaussian Spectral Decomposition (GSD) framework that leverages modeled spectral end-members and spectral-geometry analysis to decompose remote sensing reflectance (Rrs, sr−1) into constituent-related abundance components. Using a validated bio-optical forward model, we generated a compact set of modeled end-members (MEMs) that capture the main spectral variability associated with phytoplankton, suspended sediments, and colored dissolved organic matter in the Dutch Wadden Sea, the Netherlands. Their spectral geometry was assessed relative to end-members extracted from in-situ hyperspectral datasets (EEMs) and simulated reflectance spectra. MEMs formed a compact and well-separated spectral simplex that consistently enclosed both measured and satellite-derived Rrs spectra, yielding low reconstruction errors. The performance of an Endmember Extraction (EE) algorithm was examined to determine its suitability for multispectral satellite retrievals. While the EE algorithm effectively identified extreme spectral signatures of in-situ hyperspectral data, its performance degraded in the satellite-derived multispectral domain due to band aggregation and atmospheric correction uncertainties, resulting in higher reconstruction RMSE and larger spectral-angle errors than MEM-based decompositions. The results demonstrated the advantages of MEMs for stable and generalizable spectral decomposition. End-member abundances derived from GSD were used subsequently as predictors in machine-learning regression models to estimate chlorophyll-a and total suspended matter concentrations. The approach yielded high predictive accuracy for both in-situ hyperspectral and Sentinel-2/3 matchups, accurately representing spatial gradients linked to tidal mixing, sediment resuspension, and phytoplankton dynamics. The findings establish GSD as an interpretable, sensor-agnostic, and operationally scalable framework for coastal water quality monitoring and provide a pathway for extending spectral decomposition methods to broader aquatic environments.